Migration, Transfers and Labor supply: Evidence from Haiti

Transcription

1 Migration, Transfers and Labor supply: Evidence from Haiti Barrera Oscar David * and Zanuso Claire ** * Paris School of Economics ** PSL, Université Paris-Dauphine, UMR DIAL December 11, 2017 Abstract The impact of migration on both sending and receiving countries has long been analyzed. Yet, little is known about the labor market impacts of remittances, despite their magnitude in countries with considerable emigration. Do remittances affect the households labor supply in developing countries? This paper investigates the heterogeneous effect of transfers and migration on the household labor decisions in Haiti in We use the Post-Earthquake Living Conditions Survey (ECVMAS) conducted in late 2012, the first nationwide socioeconomic survey to be taken since the earthquake. The empirical analysis of this paper is based on the NELM framework theory which attempt to explain the effects of transfers and migration on labor force. In order to address a suitable identification strategy, we present a probabilistic model where outcomes are the decision of participation and ratio of household in the labor market, and a IV-tobit specification for the number of working hours. We use the effect of the 2010 earthquake as an instrument for migration. We show that the more affected the household were by the earthquake (PGA), the less they are likely to participate in migration. The results show that participating in migration increase the probability to receive transfers by 10%. Migration affects significantly household decisions however such results are not observed at individual level. Our results also suggest that the effect depends on the nature of transfers (external, internal, in cash, in kind and frequency) and the individual s position in the household (head or secondary members). We find significantly lower propensity to work for individuals in recipient households and negative effect of transfers on the aggregated hours worked in the household. Keywords : Labor Supply, Remittances, Natural Disasters, Haiti JEL Classification: F24, F22; D1; J22 ; Q54 We would like to thank participants at the 6th PhD Conference in Applied Economics (Turin, Italy, December 2013), 31ème Journées de la Microéconomie Appliquée (Clermont-Ferrand, June 2014), 63ème AFSE (Lyon, June 2014), International Conference 2014 of the Courant Research Centre Poverty, Equity and Growth in Developing and Transition Countries (Göttingen, Germany, July 2014), 26th European Association of Labour Economists (Ljubljana, Slovenia, September 2014), 3ème Atelier DIAL-GREThA Développement (Bordeaux, March 2015) for their helpful suggestions. Any remaining errors are our own. 1

2 1 Introduction Remittances to developing countries were estimated at $404 billion in 2013 (World Bank Migration and Development Brief 2014), sparking interest and debate among researchers and policy makers. At the macroeconomic level, remittances can either boost the economies of receiving countries or create a culture of dependency and thus slow down economic growth. At microeconomic level, remittances have the potential to improve the well-being of family members left, but can also lower the labour force participation. Haiti ranks in the top 10 remittance recipients in the world, with remittances having accounted for more than 21% of GDP in However, if 30% of Haitian households declared receiving external transfers in cash, the importance of private transfers in Haiti goes beyond remittances, since the same proportion reported receiving internal transfers in cash. Overall, around 70% reported receiving at least one type of transfers, internal or external, in cash or in kind. Moreover, Haiti is one of the top 10 emigration countries and top 5 for tertiary-educated emigration (almost 85% in 2000 according to the World Bank Migration and Remittances Factbook, 2011). However, in broader terms, the unemployment and informality are disproportionately high compared to other countries at similar level of GDP per capita. For example, our data set suggests that around 93% of the workforce is informal 1 After the earthquake in 2010 there was a massive spike in emigration from Haiti, especially to the USA, Canada and, to a lesser extent, to France. This phenomenon was in fact supported by a set of government policies that sought to boost emigration as a way for the economy to recover after the natural disaster. One policy was to provide special status to migrant workers from Haiti in the USA which led to a dramatic increase in the outflow of human capital from the island 2. But what is the real social effect of these kind of policies? Has the outflow of workers after the earthquake led to an increase in Haiti s social welfare? Has the loss of human capital due to migration been compensated by the additional inflow of remittances? We seek to provide 1 The ILO (2003) defines informal work as employment that does not provide coverage in terms of the institutional social protection scheme. Workers in this sector also include unprotected employees in the formal sector. In this paper informal employment will be defined as all contributing family workers who are independent workers in the informal sector and all employees without written contracts and who are not benefiting from social protection. De Vreyer and Roubaud (2013) 2 The US government introduced an exceptional visa type for Haitian workers 2

3 some insights to these questions by assessing what is the role of migration and transfers on labor supply in Haiti. In this paper We investigate the changes induced by the flows of transfers and migration on the the labor status of left behind household members, comparing them to non recipient households. We find a positive and significant causal effect of transfers and migration on labor supply at household level. We also show that the effects are heterogeneous and depend on the transfers frequency, on whether the transfers are in cash or in kind and on whether the transfers are internal or external. We also highlight that recipient households are substantially different from non recipient households in terms of welfare, consumption and level of income and level of education of the household head. The rapid increase in migration after the earthquake has created concerns among policy makers about the way some important sectors of the Haitian economy might be affected due to the massive migration outflow. For example, 40.53% of rural households have at least one migrant member and rural migrants represent almost 45% of total migration after Migration in rural areas might have a negative effect on food production and crop income, potentially affecting food security in Haiti. Additionally, if the income of rural households is more sensitive to migration than the one of urban households, then policies oriented towards increasing migration might contribute to increase the income gap between urban and rural areas. Another effect of transfers that we address in this paper is its effect on informal workers and, more generally, on the informal sector. It is hypothesized that workers in the informal sector tend to face more credit constraints than individuals in the formal sector. Therefore, recipient households might use transfers in order to loosen the credit constraint and develop their own business. Thus one could expect that, conditional on migration, transfers might increase the labour participation of individuals in the informal sector. Heterogeneous responses, including frequency of transfers, origin of the transfers and type of the transfers (kind of cash) will be addressed as well. In order to address the identification challenges, we will use a probabilistic model. The outcomes variable is the decision of participating in the labor Market. An IV-Tobit specification models the number of working hours of headship and aggregated working hours in the household. In 3

4 order to account for the endogeneity of the variables of interest (migration and remittances), our identification strategy relies on the earthquake effects on migration and the average transfers at commune level, as widely suggested by the literature. Despite the importance of internal and external transfers in Haiti, there is few empirical evidence about the impact on labor outcomes, partly because of the lack of reliable micro data. Jadotte (2009) aims to estimate the impact of remittances on household labor force, using migration as an instrument of transfers which is not completely satisfying. 3 This study provides several original contributions to the literature. It is the first study that considers the effect of the earthquake on migration, transfers and labor outcomes. Moreover, we take into account the heterogeneous effects of transfers on labor decisions, whereas, to our knowledge, no other papers have included a differential analysis on the frequency of transfers. Furthermore, we apply nonconventional econometric tools to the empirical identification method. The outcomes of this sort of investigation have important implications for immigration policy in net recipient countries as well as the analysis of the economic impact of migration. We discuss these issues in the conclusions. 1.1 The New Economics of labor Migration (NELM) The fundamentals of the New Economics of labour Migration is presented in Stark and Bloom (1985). NELM is centrally based on the idea that migration decision is taking place within a large context, mainly household influence, which potentially consists of individuals with diverse preferences and differential access to income and is influenced by its social environment. Households prefer to make this decision collectively not only to maximize income, but also to minimize risk and loosen constraints created by a variety of market failures, such as the lack of insurance system and labour market failures. Finally, relevant to this study is that NELM is taking into consideration that the effect of income on utility may not be the same for a given individual or household across socioeconomic backgrounds. The high incidence of informality in Haiti leads to families creating different mechanisms to 3 A central point of NELM framework is that even if migrants did not send a transfer to the household, the household would still have lost one member who could have increased the production of the household, therefore other members of the household would increase their labor participation. 4

5 increase their own production (mainly concentrated in the informal sector). Thus, in order to increase income, urban and rural households might invest in a new technology to make the transition from familial to commercial production. However, these households, in most of the cases, do not have access to either credit or income insurance. Therefore, by sending a family member abroad to work, households might see this emigrant as an investment with future returns. Thus, once migrants become established in their destinations, they send back transfers that might work as insurance. However, for some households, in particular rural ones, the initial cost of sending migrants can be particularly high. Let us consider a household with two possible production activities (e.g. a low and a high return activity). A household may invest its fixed resource T 4, such as as land or a building in which to produce a manufactured good in eitherlow return (e.g. grain) or high return (e.g. manufacturing) activity. Let Q i for i = 0, 1 denote output from these two activities, respectively. A set of household characteristics, Z y, explains the returns from investing household resources in each activity (these characterstics may be education, age, sex, location, number of children etc.). Let us also include a level of effort e that the household has to make in order to reach a given level of Q i by keeping other factors constant. Then e in the model is highly correlated to T. At relative price p 1 p 2 the household will specialize in the high return activity, Q 1. Its output will be: Q = f 1 (T (e ), Z y ) and its income will be Y = g(q(e) ), where e is the optimal level of effort that might lead to Q However, the household may face market constraints to invest in the high return activity: c(.) = T 1 (e 1 ), with c(.) implying that the household can invest until T 1 (e 1 ) 5 is less than T (e ) and e 1 < e. For 4 This is an adaptation of the model presented by (Stark and Bloom, 1985). We use our own notation based on Taylor et al. (2003) 5 For example they cannot buy fix capital or suitable equipment which might increase productivity then affect Q 5

6 instance, in the case of households in Haiti, households might face credit or liquid constraints c(.) which prevent households from borrowing to increase the purchase of inputs and produce Q 1 in its land or more commonly home based small business. Thus, in this example T 1 (e) is the part of household s land or capital investment of the home based business that is actually invested in the high return activity. Notice that T increases as household puts more effort in the activity e. However, at the same time, given that T (e) the household will make the decision of e as they know T 1 therefore e is higher than e. Migration plays an important role overcoming the constraints of households. Constrained families might sent migrants out to work in a wage earning job, thus, migrants could help relax the household s liquidity constraint by sending back remittances, R. Additionally the flow of remittances might affect the nature of uncertainty faced by households 6. However households sending migrants abroad have also to face a cost for it. For example, if rural households also face a missing or imperfect labour market and therefore must rely on family labour, migration may constrain the household from moving toward the high return activity. Therefore, in order to keep same level of income, individuals might choose to increase e. Hence, NELM hypothesizes that the constraint limiting the amount of the fixed resource that can be allocated to the production of the high return activity, T 1 is a function of migration and remittances: c(r, M) = T 1 (e 1 ). If c c > 0 and < 0, since migration leads to a reduction in family labour supply and a rise R M in available capital for production in the source of household, we could expect migration to lead to an increase in the labour participation of rural and informal workers; as for remittances, we expect households that are receiving remittances to replace credit constraint will increase their participation. Thus, constrained output in the high return activity is Q c 1 = f 1(T 1, Z y, e i ), and in the low return activity it is Q c 0 = f(t T, Z y, e i ). Constrained household income, Y c, is given by: 6 We are considering in this study the direct impact of transfers on labour participation in Haitian households, we do not attempt to address issues related to risk. 6

7 Y c = g(q c 1(e c 1), Q c 0(e c 0)) where Y c < Y is the unconstrained outcome. It is important to note that e R and e are unknown. Thus the overall effect of migration M on household s labour participation and income is ambiguous. However if the level of effort in a given household is function of human and financial capital in the household, the impact is not likely to be zero. Therefore finding that migration or transfers significantly affect any nonmigration source of income (hence level of effort) in the household sending migrant or household receiving transfers would support the NELM theory. Of course, the sign of an activity-specific migration effect, like the total effect of migration on effort, is indeterminate a priori. This theory suggests that migration might lead to increase effort in the high return activity (Q 1 ) if it release the constraints, c(.) which is a limitation for that activity. However, the behavior of household might differ in other activity (Q 0 ). For example, if one individual of the household in a low return activity leaves, stayers would not have much incentive to increase labour participation. To our knowledge, there is no empirical application of the NELM theory to Haiti. Taylor et al. (2003) 7 test NELM for rural areas in China. They attempted to identify separately the effect of migration and remittances on different types of income in rural households of China. Similarly Benjamin and Brandt (2002) found evidence that off-farm labour market participation loosen risk constraint on household farm activities in China. If migrants play the role of financial intermediate among informal workers of Haiti, as this study suggests, the ex ante incentive to migrate must be relevant. Propensity to migrate might depend then on the capacity that the households have to fund investment or whether the loss of labour to migration carries significant costs in terms of forgone revenues to the household. 7 Further works are Taylor (1992), Lucas (1987) and Taylor and Wyatt (1996). 7

8 2 Literature review and evaluation of migration theories Despite the increasing amount of literature on migration, remittances and labour market, the empirical evidence in the in the Latin America and Caribbean region is still limited. Amuedo- Dorantes and Pozo (2006) found that remittances led to a reduction in the hours worked in all sectors, therefore the income effect dominates substitution effect in female recipient households in Mexico. However, Cox-Edwards and Rodriguez-Oreggia (2009) did not find evidence of labour force participation effects of persistent remittances for the same country. By contrast, Acosta (2006) found that labour supply for men remains unaffected by the increase in remittances in El Salvador. In another important study, Funkhouser (1992) found that remittances reduce labour force participation in Nicaragua -Managua- whereas slightly increasing informality. Medvedev (2008) found similar conclusions for Jamaica. Meanwhile Itzigsohn (1995) observed a negative correlation between remittance recipient and employment status, addressing in his analysis the issue of endogeneity due to an hypothetical reverse causality. Further literature suggests also that remittances are negatively correlated with changes in other sources of income at the household level, meaning that they might act as an insurance for households in case of an unexpected shock. Interestingly, Clarke and Wallsten (2003) concluded that remittances do act as insurance, but only partially. They show that remittances increased by about 25 cents for every dollar of damage the hurricane inflicted on the household. Negative correlation between changes in remittances and changes in income from other sources is consistent with the hypothesis that remittances increase in response to reductions in other sources of income. This is also consistent with the alternate hypothesis that household members change their behavior when they receive remittances. More recently Bertoli and Marchetta (2014), by using a matching method and using data from surveys, analyzed the influence of the recent wave of migration on the incidence of poverty among stayers in Ecuador. They found that Ecuadorian migration reduced the incidence of poverty among migrant household. Given that our study is embedded in the particular context of the earthquake that stuck Haiti on January 12, 2010, we also take into consideration the literature addressing the interlinkages between natural disaster, migration, transfers and labour markets. For example Clarke and 8

9 Wallsten (2003) found that in Jamaica remittances increased when the household was hit by an exogenous shock. However the main source of controversy about this paper is regarding the method to measure the disaster. A referential work for this paper is the paper written by Gagnon (2013) who posited that the changes in labour outcomes and remittances observed at the household level in Honduras are directly linked with the impact of Hurricane Mitch. His results suggest that households that were more affected by the hurricane Mitch in Honduras were more likely to send out migrants, although the impact was not felt until the medium run. He also found a spillover effect of remittances on other municipalities over time. With a different approach to the effect of natural disasters, Becchetti et al. (2012) uses experimental data, and finds that households that experienced the earthquake are more altruistic. Furthermore, Mohapatra et al. (2012) work in a cross country model, assuming that there is no heterogeneity on the effects. Interestingly their data set of disaster was obtained from Center for Research on the Epidemiology of Diseases (CRED), International Emergency Disasters Database (EM-DAT). The main critique raised against this paper is focused mainly on the methodology used to create a variable for disasters, since they aggregate all kind of disasters: meteorological, climatological and geophysical disasters which indeed might have different effect depending on the circumstances. This paper provides evidence of different use of remittances within developing countries after a natural disaster. For instance, in Bangladesh, they show that per capita consumption was higher in remittance-receiving households than in non-receiving ones after the 1998 flood. They also found that Ethiopian remittance-dependent households seem to use savings rather than sell livestock to cope with drought. Additional findings show that in Burkina Faso and Ghana, international remittance-receiving households, especially those receiving remittances from high-income developed countries, tend to have housing built of concrete rather than mud and greater access to communication equipment, suggesting that they are better prepared to face natural disasters. Addressing issues associating poverty and inequality with remittances, Lamaute-Brisson (2002) concluded that when remittances are targeted to financing economics activities, they are devoted to informal operations in the service sector with low productivity. Additionally she found that, 9

10 even though remittances allow some households to escape poverty, there is no evidence that they reduce inequality since these remittances are concentrated in the top income distribution deciles. Jadotte (2009) estimated the effect of remittances on labour supply for Haiti, using the Living Conditions Survey of He showed that that there is a negative impact of remittances on labour market participation using both an IV Tobit for hours worked and an IV Tobit for labour market participation. He brought evidence that women s wages might be more sensitive to remittances flows than men s. 3 Empirical strategy This paper aim at studying whether receiving transfers affects household labour supply, by estimating the following equation: L i = α + β log(t h ) + δ Mig + γx ih + ε i (1) where L i is the labour supply outcome. The variable of interest, log(t h ), is the log of transfers in Gourdes -HTG- received during the year of reference by household h. X ih a vector of individual i and h household characteristics. The vector X includes variables at regional level such as zone, and regional unemployment rate; variables at household level such as number of adults in the household, consumption, number of injured or disabled people on the household, whether the household is owner of the dwelling, number of children in the house, welfare index, share of individuals in unemployment and variables at individual level, age, sex, education level 8. The coefficients of interest are β and δ. Thus, if due to the increase in remittances, the reservation wages of households members increases, it would be expected that we will have a β with negative sign. On the other hand it might be expected a positive sign of δ, as observed in the referent framework NELM. Household sending out migrants might experience a reduction in household production; therefore remaining household members might enter the labour force or increase the labour supply to compensate for the loss of income. However, despite including a wide set of controls (vector X ih ), serious problems remain with obtaining an unbiased estimate of β. This 8 We will not present the coefficients of all these variables 10

11 issue will be addressed in a subsequent section of this paper. The null hypothesis associated with NELM are that neither migration, Mig, nor transfers log(t h ) affect labour, that is β, δ = 0. However, not all households that sent out migrants receive transfers afterwards, the inflow of transfers is the result of sending out labour migration. Thus, controlling for migration, transfers are affected by household and commune characteristics X h which, in turn, affect the success of migrants or stimulates migrants to send remittances. Thus we model the first stage of transfers as follows: log(t h ) = η 0 + η 1 Z Mig + η 3 X ihc + ν i (2) It is also performed migration as a function of household and commune characteristics Mig = g(φ; X ihc ) + v i (3) In order to identify the causal effect of migration and transfers on labour we will use an instrumental variable (IV) approach, where equations 2 and 3 will be estimated by using exogenous variables which will be discussed in subsequent sections Variables Outcomes We will analyze six labour outcomes. First, L 1i will be a dummy variable taking the value 1 if the individual participates in the labour market (employed or unemployed) and 0 otherwise. This approach will allow us to identify directly whether remittances affect the labour decision of individuals of working age, but does not allow us to infer about the status of the worker (formal and informal job) due to the fact that the formal status is unobserved for unemployed individuals 10. The second outcome, L 2i, is a binary variable taking 1 if the individual is actually working and zero otherwise. Contrasting these two outcomes is interesting to understand to what extent transfers and migration not only affect the labour decision but also whether transfers or remittances have a significant effect on the unemployment rate. 9 This strategy is based on the empirical model proposed by Taylor et al. (2003) 10 All models will be run on individuals older than 12, since for developing countries the ILO defines as children those individuals younger than 12 years. 11

12 We also suggest focusing only on those employed individuals who are working. Thereby, the outcome variables number three, L 3i, is the total number of hours that the individual devoted to work during the week of reference. Additionally, the fourth L 4h is a variable at household level which aggregates all working hours in a primary job of all members of the household. Finally L 5i is the total number of hours (primary + secondary) of the individual i in the household h. A Tobit approach is a suitable strategy of identification due to the large number of observations with zero working hours in the predicted variable, the model is censored in 0 and 120. Thus it will be possible to analyze the behavior of informal workers and workers in the informal sector. On the one hand, a negative effect of remittances on the outcome is expected for those individuals in the formal sector. On the other hand, a positive effect of remittances on number of hours worked is expected among informal workers. 3.2 Sources of bias Some econometric issues arise in the estimation of equation 1. First, L 1 and L 2 are binary variables and L 3 L 4 and L 5 are zero inflated continuous variables. Therefore the estimation of the equation 1 using an OLS approach would yield biased and inconsistent estimates of the impact of remittances on the probability of participating in the labour market, getting a job and the number of working hours. Second, the transfers and error term in equation 1 may be correlated. Transfers might be endogenous, therefore the coefficient estimated might be biased. For instance, some unobservable household characteristics, such as the entrepreneurial spirit of the household head is likely to jointly determine domestic income and remittances. Additionally, endogeneity of transfers can be due to omitted variables bias which may exist if transfers are related to wealth issues which, in turn, might be associated to the choice of work hours by the household member. Another important sources of bias affecting the identification is reverse causality. Households might either make their labour participation decision depending on the level of remittances received, or ask for financial support to members of the household that have migrated because of their current labour status. To account for the fact that the variable of interest is binary and the endogeneity of the inde- 12

13 pendent variable, we use in a first approach the model used by Jadotte (2009), an IV Probit. We also estimated the Probit model for comparability purposes. For the continuous variables L 3 L 3 and L 3 we use Amemiy Generalized Keast Squeres (AGLS) estimator for the Tobit with endogenous regressor IV-tobit described by Newey (1987) cited in Amuedo-Dorantes and Pozo (2006). We estimate a first stage remittance and migration equation by using two additional exogenous variable instruments discussed in a subsequent section. 3.3 Selectivity A serious concern refers to the potential non-randomness of the treatment group. Abramitzky et al. (2012) concluded that during the mass migration era from Europe to the USA in the early 1900, Norwegian migrants from urban areas were negatively selected from the sending population, whereas rural migrants were generally positively self-selected, hence their likelihood of success was higher. Jadotte (2009) also found that the average income of household recipients of remittances is higher due to a selection bias among migrants in Haiti (based on the 2001 households survey). In order to confirm selectiveness in remittances in this survey, we constructed a welfare index at household level. Factor Analysis approach 11, developed by Lawley and Maxwell (1962) is used to estimate the welfare index. The index includes the most relevant social characteristics of each household (quality of the dwelling, access to drinking water and number of rooms in the house). We thus attempt to identify sources of selectivity in order to control for it during the identification stage. Based on ECVMAS 2012 data, the first graph of the figure 1 confirms that self selection of migrants reported by Jadotte (2009) remains constant since We notice that the labour income is higher among recipient households. It is also confirmed that these distributions are significantly different one to another at 1% level of confidence by using the Kolmogorov - Smirnov test. This result suggests a clear case of self selection. We found that recipient households in 2012 have performed better in the index. The kernel densities on the centiles of the welfare index confirms that recipient households are more likely to have higher level of welfare Based on Sahn and Stifel (2000), who constructed a wealth index based on the factor analysis approach 12 We also test the difference in these distributions with the Kolgomorov-Smirnov. This test reports the 13

14 Figure 1: Kernel distribution of working income and welfare index by recipient status Source: ECVMAS, 2012, IHSI / DIAL / World Bank; Authors construction 14

15 3.4 Source of bias: local economic performance Another source of bias might be that the labour market decision might depend on the economic environment in the region and, at the same time, remittances might be jointly affected by the economic and social context of the area. This is a major issue in most of the studies which have attempted to address similar questions in other contexts. For instance, Acosta (2006) found that labour participation of men remains unaffected by the increase in remittances in El Salvador but its econometric approach does not include control for economic performance. The instruments suggested by Acosta (2006) are the migrant networks at the village level and the number of international migrants who returned two or more years ago, as well as the propensity to migrate. However, other relevant social conditions of the geographical variables might have an important role in the migration decision. Some region or commune characteristics might affect the decision of migrate (e.g. availability and quality of public services, violence, pollution). In order to account for the potential effect that the social context might have upon the decision of migrate we propose to control our structural model by the regional unemployment rate. Additionally, labour participation might also be explained in part by local conditions. A large body of literature has attempted to identify the effect of social atmosphere and migration decision. A number of studies have also shown the impact of unemployment rate on the decision to migrate. For instance, Di Pietro (2005) found that the regional unemployment rate in Italy lead to increases in regional migration among recent Italian graduates. Therefore the local unemployment rate should be included in the specification as a control. The first graph in figure 2 illustrates that unemployment status is highly correlated with the level of received transfers in the country during Interestingly we notice that unemployed people tend to receive more transfers. THe second graph suggests that a number of unemployed people are concentrated in areas where unemployment is higher. Thus, one unemployed person who is eligible for employment and is able to work might could decide to stay at home because she does not find suitable employment options or she was met with lack of success when she maximum difference between the two cumulative distributions, and calculates a P-value. The null hypothesis is that both groups were sampled from populations with identical distributions. The populations may differ in median, variability or the shape of the distribution 15

16 Figure 2: Kernel distribution of unemployment rate and log of Transfers by labour status Source: ECVMAS, 2012, IHSI / DIAL / World Bank; Authors construction applied. Therefore, discouraged workers might be the result of the high unemployment rate in the region. This implies that regardless of transfers or reservation wage, a higher unemployment rate in a given region might dissuade people from working. 3.5 Heterogeneous effects Frequency We also include dummies for transfers frequency: high and low frequency. Those households that have received more than 4 transactions during the reference year are considered as household receiving transfers at a high frequency and, between 1 and 3, low frequency 13. In order to identify the heterogeneous effect of frequency between recipient households, we include an interaction term for high frequency. It is expected to have a negative and significant effect of highly frequent transfers in cash on labour participation. This implies that for an increase in the amount transfers in cash highly frequent by 1% we expect a reduction of the likelihood of participating by γ 1 % and a reduction of the number of working hours in γ 2 % (see equation 1). The table 1 presents the mean and share of households receiving cash transfers at high and low frequency differentiated by whether the transfers are internal or external. As expected, we notice that external transfers are, on average, larger than internal ones, because workers abroad might have better working conditions, and thus they would be able to send higher transfers. It 13 Frequency is only considered for transfers in cash since the survey does not provide information about the frequency of transfers in kind. 16

17 is also clear that the effect of high frequency transfers is on average 6 times the low frequency effect and this is also an expected result and consistent with the literature. Table 1: Frequency of transfer by origin Origin of the transfer External Internal Total A. High Frequency: Mean 3, , , Standard D 3, , , Population 257, , , 542 Share of pop Sample B. Low Frequency: Mean Standard D Population 392, , ,218 Share of pop Sample C. Total: Mean 1, , Standard D 2, , , Population 650, ,373 1,168,760 Share of pop Sample 1,348 1,352 2,510 Source: Author s computation from data in ECVMAS Within household members In an idealistic scenario we would measure how remittances affect labour decision of every single adult in the household. However, it would require precise information in two aspects. First, the labour status of each individual within the household and second, precise information on the intra-household allocation of remittances. As for the former, the survey provides wide information; therefore, we identify those individual employed, unemployed and those not participating on the labour market. Nevertheless, for the latter, the survey reports the amount of transfers that each household received during the referent year, but does not provide any information about intra-household allocation. We test the correlation between remittances received by the household h with the labour participation of the head of the household, the spouse and secondary members, the spouse (figure 3). The first graph in the figure suggests the correlation between transfers received and number of weekly hours worked the household head. The second graph portrays the correlation of received transfers and spouse s working hours. The third graph reports the correlation between the labour participation of the secondary members of the household and the received transfers. 17

18 Figure 3: Plot of responses of total transfers on individual labour within the household Source: ECVMAS, 2012, IHSI / DIAL / World Bank; Authors construction The result confirm that transfers have heterogeneous effect on intra-household labour decisions. Therefore, the estimation will be run at individual level including equal value of remittances for every single adult in age of working in the household. Including different dummies for each adult in the household will thus capture the effect of the remittances within the same household Endogeneity of transfers and migration We have already described the main shortcomings that make difficult the identification of the causal effect of migration and transfers on labour participation. In order to address the endogeneity problem in transfer and migration this paper will rely on the use of an instrumental variables approach. Earlier we presented the theoretical framework followed in this paper, suggesting that migrant remittances may be influenced by village norms to remit. Therefore, following Taylor et al. (2003) we use the average level of remittances among families in the village, dropping the observed household, as a proxy for village norm. The main assumption of this approach is that the village norms regarding transfers affect each household s remittance level but it has no effect directly on the labour decision. One potential critique of this assumption is that transfers might have a direct effect on local labour demand. For example, households in communes receiving a higher amount of transfers might have more consumption capacity; therefore the aggregated demand of these communes might be higher, thereby possibly increasing labour demand or wages. However studies such as Bussolo and Medvedev (2007) suggest that higher levels of transfers in Jamaica have general equilibrium effects only in the midterm. Moreover one can think that high level of remittances could lead to an increase of inflation when productivity remains constant, therefore real wages will remain constant and won t induce any effect on labour decisions. This 18

19 hypothesis is confirmed by checking the correlation between the instrument and the individual labour supply(0.0350). In order to guarantee the exogeneity of this variable, Taylor proposes to use the average transfers at commune level excluding the observation of the household as a proxy for village norms. Table2 confirms the significance of the instrument. We also include the conventional set of controls and some additional relevant variables in order to account for differences related to the frequency of received transfer, type of transfer (kind of cash) and origin of the transfer. The first stage of the structural model estimates transfers. The equation of transfers is estimated using the specification of migration from table 2 and employing the instrument of average transfers received in the commune. Column five predicts the probabilistic model of remittances where the dependent variables takes value of 1 if the household h of individual i received any kind of transfers, 0 otherwise. Column six runs the total amount received at household level on the same variables, high significance of the instrument and R 2 of 0.34 and 0.43 suggest a well predicted model for transfers. In order to attend to the endogeneity of migration, the total number of migrants at household level is estimated. Thus a migration process model is estimated in a first stage and predict total number of migrants per household 14. The outcome variable (migration) is analyzed to see if the migration decision might partly be explained by the 2010 earthquake and the network of migration before the earthquake 15. We then will instrument migration using the earthquake effect on Haiti in In 2010 Haiti experienced an earthquake that definitely affected the general equilibrium of the economy and thereby affected household decisions regarding its labour participation. A natural disaster is usually a good source of exogenous variability that offers a suitable instrument to correct endogeneity. However, in order to instrument migration with the effect of the earthquake we will make two assumptions. First, the earthquake does not have a direct effect on labour participation. Instead, it might affect the household decision through migration. However, it is likely that the exogeneity assumption does not hold in this case and two main channels have been identified, the first one being economic performance. It turns out that after the earthquake more affected areas could have 14 We are adding here both migrants that went out before and after the earthquake 15 This variables is computed by crossing information of current level of remittances and migration after A significant limitation at this stage is the lack of information of those migrants that did no reported transfers before the earthquake 19

20 Figure 4: Total number of migrants per household Figure 5: Author s calculations based on the ECVMAS. received wider attention by either the government, NGOs or other donors; as a result, in the mid term, such areas could have experienced higher growth than others. By contrast, the earthquake could have affected the local production via capital destruction, leading, in turn, to lower demand of local labour. Thus in order to control for these channels, we include a control of regional unemployment rate. The second channel through which the earthquake can affect labour supply is remittances. Indeed some dwellings could have been more affected by the earthquake. Thereby, constrained households, with migrants living abroad, could have found in remittances a suitable way to invest in the reconstruction of the dwelling. Hence it is expected to find a positive correlation between remittances and damage in the household. However, some households might decide to get a loan in order to set up the reconstruction of the dwelling, thereby, depending on the degree of destruction, some households in 2012 might remain in debt. Therefore we are controlling our structural model by the debt status of the household head. 3.6 Predicted migration and transfers By looking at the figure 4 one observes that the data are skewed to the right, so clearly an OLS regression would be inappropriated. This sort of data usually follows a Poisson distribution, 20

21 therefore some type of Poisson analysis might be suitable for this approach. The first issue to be addressed is the potential overdispersion of the data since the number of migrants per household vary from 0 to 11. From statistical theory it is inferred that in a Poisson distribution the mean and variance are the same. Our data, however, suggests that the variance of migration by household is just twice as large as the mean. The distribution thus does not display overdispersion 16. The predominant feature of the figure 4 is the strong concentration of observations at zero in an otherwise smooth distribution. According to the graph 4 three out of four households in Haiti decide not to participate in migration. The histogram of migration suggests that after the hurdle is crossed, those households deciding to send one or more have a decay trend that is relatively smooth. The distribution of migration leads us to believe that the statistical model determining the binary outcome of the number of migrants as zero or one differs from that determining the total number of migrants. In the conventional counting data models, like Poisson, these two processes are forced to be identical. Thus in this paper, instead of the standard Poisson model, a model that allows the binary choice process to differ from the frequency estimation is used. The model proposed by Bohara and Krieg (1996) to predict migration suggests that the migration decision distinguishes persons in an important way. Thus they state that person who uproot themselves and possibly other family members by migrating are different from those unwilling to undertake such a venture. Thereby the empirical model for migration is separated into two steps. The fist step is to estimate whether households decide to participate in migration. Thus, conditional on the migration decision, the frequency of total migrants in 2012 is estimated 17. It seems natural to allow the statistical model governing the binary outcome (participating into migration or not) to differ from that which determines the frequency of migration.the hurdle model defined by Cameron and Trivedi (2013) is performed to estimate migration. Mullahy 16 In order to confirm this claim, a poisson regression is run in order to test the poisson goodness-of-fit Deviance goodness of fit = with P rob > chi2(1150) = Additionally Pearson goodness of fit is performed: with P rob > chi2(1150) = The lower value for chi-square of both goodness of fit: deviance and Pearson, are another indicator that the poisson distribution is a good choice. The insignificance values of the test statistic from the goodness of fit indicates that the model the variable outcome does not suffer of overdispersion. 17 We are including both external and internal migrants. Differences in behavior between them will be addressed in the estimation 21

22 (1986) suggests the following formalization 18. This model hypothesis that the decision is governed by the binary outcome of whether the household send a migrant abroad or not. Secondly, when the household decides to send out a member, then the number of additional members sent within a given time frame is allowed to follow a truncated-at-zero Poisson model. Therefore when a household decide to participate into migration it implies a positive realization, known as crossing the hurdle. The joint likelihood function for the hurdle model Poisson model is: L = Π n t=1 [P r(y 1t = 0)] y 1t [1 P r(y 1t = 0)] 1 y 1t [P r(y 2t y 2t > 0)] y 1t (4) Where y 1t refers to a dummy variable taking value of 1 if the household participates in migration, zero otherwise and y 2t is the frequency of migration. The first two terms represent the binary outcome representing migration decision and the last term captures the conditional Poisson function. The following expression ilustrates the joint likelihood function L = Π n t=1 [F (X it β 1 ] y 1t [1 F (X it β 1 ] 1 y 1t [λ y 2t /((exp(λ t ) 1)y 2t!] y 1t (5) Thus, we assume that the decision whether or not an household participate in migration is represented by a probit function F(.). The set of independent variables X 1t affects the migration decision. The frequency of migration is influenced by X 2t through a link function λ 2t = exp(x 2t β 2 ) which ensures that the expected frequencies are positive. The hurdle Poisson model is also suitable for taking into account the overdispersion or underdispersion of data (Mullahy, 1986). Thus we can omit alternative approaches such as negative binomial. Following Bohara and Krieg (1996) approach the following variables play an important role in the migration decision: education of migrant and head of household, income level in the household, labour status of other members in the household, age of the household members, gender, ownership of the house, education level, disability and number of children in the household. By following the approach of Taylor et al. (2003) a set of variables at commune level will also be included in the model. The level of damage of the earthquake at the household level will be included as control. 18 We are following the notation proposed by Allen Mc Dowell in the Stata Journal 22

23 Thus we will estimate the equation of migration using the Hurdle equations approach. Josepth Hilbe (N.D) created a suitable algorithm in Stata to run Hurdle by including a logit probabilistic model in the first stage; however McDowell (2003), by simulating data on migration, proved that better predictions are reached by using a binomial probability model with the Stata s cloglog command. Then manually they estimate the parameters of the truncated Poisson model with the command trpois0 developed by (Hilbe, 1999)). The results are presented in tables 2 columns 1, 2, 3 and 4. Two instruments to identify migration were used. The first one, following Taylor et al. (2003) approach, was the percentage of the workforce out-migrating after 2010 at commune level. As for the second one, we will use PGA (Peek Growth Acceleration) as an objective measure of the intensity of the earthquake at commune level. A positive and significant effect of PGA on migration is expected. The correlation of the predicted value of total migration and the actual variable is 0.75 by using the hurdle approach method. An additional issue for estimating the migration decision model is the identification of the migration process. Empirical evidence suggests that one important factor in the decision of migrating is the social network which reduces settlement costs (expenses associated with migration ) and therefore makes the financial travel less constraining. Additionally, having contact with individuals in the host country makes it more probable to find a job faster. Migration rate at commune level is used as network variable for the identification of the migration equation. Migration share at commune level is computed by dividing the number of migrants in a commune by the total population in the same commune. In order to ensure some variability, the index is interacted with the number of grownups presents in the household. As expected, migration is positively associated with transfers. Making the decision of sending a migrant out increases the probability of receiving transfers by 10 percent. However we additionally show that migration has decreasing returns to scale. Observing the column six in for migration in the first step approach 2 it is noticed that for one additional migrant the household has abroad, the transfers drop by 237 HTG (around USD 5). However households that sent out migrants for the first time can expect that migrants will contribute to the income of households in Haiti. This result is pretty much in line with NELM theory, since the first migrant represents a reduction in the human capital of the household, the household would expect to receive transfers. However once the household sends an additional migrant, the loss of human capital 23